video_OCP2026_LightCounting_CPO與NPO市場預測_20261015

So, if you think that market is getting a bit crazy, you're not the only one in this audience. And we are having fun forecasting this market. And we no longer show just the forecasts. We show how forecasts changing every 3 months. So, it's a new new line of business. I mean, it's great for selling reports. All the forecasts that presenters were showing are outdated. Now, you have to buy the new one September 26. But to be honest, I mean, we've been doing it for last 2 years. Every 3 months the forecasts keep on going up. I think only once it went down after soon after the liberation day, which ended up being completely benign. But to be honest, the latest change, you know, the difference between the blue and the red line, this is really staggering for me. I mean, we we've been increasing it up, but the the step up in the last 3 to 6 months is is really crazy. And we now forecast that the market will grow 120% in 2026. Last time we've seen more than 100% growth was year 2000. >> [snorts] >> Wait for the next slide. Let me finish with this one. So, 2027, you know, we we still forecasting moderation in the growth, but even in moderation from 120% we go to 80 70. Yes, it's moderation. It's still a crazy growth rate. Um even for the blue curve, which you see goes through kind of a flat flattish period in 29 2030, our CapEx forecast is still very optimistic. So, we are getting to the $3 billion that Sue mentioned, which may be the best case scenario. That's consistent with our blue line, but there's some effects um you know, timing of optics for scale-up in our forecast. It's 2029-2030. By 2031, it will return the market to double growth rates. We expect more improvements in efficiency and how optics is being used in scale-up. With multi-plane architectures, with higher rate of switches, we may see actually less um transceivers per GPU, probably starting in 2027 and 2028 already. And by 2029, it I may actually have have some impact. So, the CapEx is still growing, but the the portion that goes into transceivers, and I agree with Frank, it's about 5% of CapEx goes to transceivers. By 2029-2030, it's going to be 4%, maybe even 3.5%. So, it doesn't have to kind of scale linearly because people find ways to use um hardware more efficiently. So, another way we been changing our reports, we no longer forecast just a single scenario. So, the most likely scenario, this blue line, is is very conservative to what could potentially happen. You know, there is a fantastic growth scenario, which actually still implies moderation in the growth rates. So, last month we made a statement that we are probably at the peak of this cycle. And what we meant in terms of the growth rate, we are clearly in the peak. So, all the three scenarios, including the green, assume that, you know, the growth rates are going to moderate. If the growth rate remains the same, this green curve would be exponential. If it's linear, it means the growth rate is moderating. It's still double digits by 2031. The red scenario, the bumpy ride scenario, is the only one that consistent with the history of our industry, unfortunately. So, one of our analyst is a mathematician by training and he runs our models and he argues every year that the red forecast is the only the right is the only one we should publish. Everything else does not conform to the history of the industry. But, being a human, you know, we have to we have we have to find a compromise. We cannot just publish a forecast for a declining market even so we know it will happen at some point. We don't know when, we don't know how sharp it's going to be, but some kind of recalibration is needed and to be honest, the sooner we see recalibration the better because we don't want to grow like crazy for another 3 years and then crash. I would rather have a moderate um recalibration now, next year, the sooner the better. And then the market returns to growth. So, that's that's the beauty of the cycles. You know, we were all worried about the the downturn, but you you have to start planning for the next upturn. And that's where the NPO and CPOs are coming. So, we have a forecast that's been published in May. Um we increased it slightly in July in um >> [clears throat] >> and in September, but since it's very forward-looking forecast, you know, what's happening in '26, '27 um is not um directly impacting the the longer term forecast. And you can see a mix of applications for scale out, for scale up. We also separated the market between NPO and CPO. And if you look at the end users, um and this is our analysis. It's not endorsed by any of this end users. But we certainly see a lot more excitement about NPO, which is a kind of a democratic version of co-packaged optics. So, we see Amazon, Alibaba, Google, Meta, Microsoft, Tencent and clearly supporting this direction, announcing products, announcing early deployments. On the other side, um you know, smaller companies like Near Clouds uh or even Oracle, which is one of the smallest in the top five US, still very large companies, many enterprise companies, they'll be very happy with CPO solutions, whether it's Broadcom or Nvidia that supply uh fully integrated systems to them. And you know, some com- some companies may deploy both. They may deploy um um NPO and CPO because there will be different applications. Uh we may we may see, you know, multiple flavors of NPO and CPOs. It's not going to be one solution fits all. Uh but certainly moving to um higher data rates per lane makes um NPO and CPO more attractive. So, if you look at the market by 2031, and that's splitting it by data rate. So, we will we will still have 800 gig ports, which are selling like crazy this year, and they'll continue to sell. With 800 gig ports, so this is 100 gig per lane technology, majority of the products will be re-timed transceivers. We We still see deployments of LPO at 800 gig, but to be honest, they are minor. They are more kind of a test exercise, a limited exercise. The same for CPO and POs. This will be proof of principal deployments a limited scale. At 1.6 terabit, this is 200 gig per lane. This will be still the bulk the bulk of the market in terms of number of ports even by 2031. And that's where a CPO NPO could easily take half of the market. Um and with 3.2 terabit ports which will just start in 2029 2030 uh but could scale in volume quite sharply. So, more than 100 million of those ports. That's where we expect CPO NPO being 75% of the market. So, that's where advantages of having optics very close to ASICs uh are the most obvious. Um So, just one more example just one more comment. I think I still have a couple of minutes. Um so, when you look at this um forecast for NPO and CPO we will be updating it next months uh including scaling applications or extended memory applications. I think when you forecast the market, you have to be aware of um emerging applications of optics. Uh first time we published forecast for CPO was in 2018. So, about 8 years ago. Uh RPE was funding projects around co-packaged optics and they reached out to us um to build a forecast and we told them for 5 years the forecast is easy to build because it's going to be zero. So, we ended up publishing developing a 10-year forecast for them. And uh I looked at it recently from curiosity. In terms of timing, we were actually fairly close. So, 2026 2027 when we expected first CPO to be deployed. But in terms of number of units, we were about 10x low on number of units. And the problem is, back in 2018, the only application we were aware of was the front-end networks, where you know, maybe only 15% of transceivers going today. So, the back-end networks, the scale-out scale-out network, that's where the bulk of the optics is going, and it's, you know, 4x more transceivers going into that application. Scale-up is another very high-volume application that we were not aware of. So, when we forecast for the future, and actually in December, we are planning to do a 10-year forecast as an exercise learning on our mistakes from 2018. So, we'd like to include new applications. So, I would like to end my my talk with a message, "Don't worry about the downturn. Plan ahead, because there going to be new applications that will make this forecast probably look ridiculously low, not even 10 years from now, you know, 5 years from now. So, let us be optimistic. Thank you.